Instrumental variables in the cost of illness featuring type 2 diabetes.

Abstract

Objective: To ascertain how an instrumental variables (IV) model can improve upon the estimates obtained from traditional cost-of-illness (COI) models that treat health conditions as predetermined.

Study setting and design: A simulation study based on observational data compares the coefficients and average marginal effects from an IV model to a traditional COI model when an unobservable confounder is introduced. The two approaches are then applied to real data, using a kinship-weighted family history as an instrument, and differences are interpreted within the context of the findings from the simulation study.

Data sources and analytic sample: The case study utilizes secondary data on type 2 diabetes mellitus (T2DM) status to examine healthcare costs attributable to the disease. The data come from Utah residents born between 1950 and 1970 with medical insurance coverage whose demographic information is contained in the Utah Population Database. Those data are linked to insurance claims from Utah's All-Payer Claims Database for the analyses.

Principal findings: The simulation confirms that estimated T2DM healthcare cost coefficients are biased when traditional COI models do not account for unobserved characteristics that influence both the risk of illness and healthcare costs. This bias can be corrected to a certain extent with instrumental variables. An IV model with a validated instrument estimates that 2014 costs for an individual age 45-64 with T2DM are 27% (95% CI: 2.9% to 51.9%) higher than those for an otherwise comparable individual who does not have T2DM.

Conclusions: Researchers studying the COI for chronic diseases should assess the possibility that traditional estimates may be subject to bias because of unobserved characteristics. Doing so may be especially important for prevention and intervention studies that turn to COI studies to assess the cost savings associated with such initiatives.